gfx/qcms/transform-sse2.c

   1 #include <emmintrin.h>
   2
   3 #include "qcmsint.h"
   4
   5 /* pre-shuffled: just load these into XMM reg instead of load-scalar/shufps sequence */
   6 #define FLOATSCALE  (float)(PRECACHE_OUTPUT_SIZE)
   7 #define CLAMPMAXVAL ( ((float) (PRECACHE_OUTPUT_SIZE - 1)) / PRECACHE_OUTPUT_SIZE )
   8 static const ALIGN float floatScaleX4[4] =
   9     { FLOATSCALE, FLOATSCALE, FLOATSCALE, FLOATSCALE};
  10 static const ALIGN float clampMaxValueX4[4] =
  11     { CLAMPMAXVAL, CLAMPMAXVAL, CLAMPMAXVAL, CLAMPMAXVAL};
  12
  13 void qcms_transform_data_rgb_out_lut_sse2(qcms_transform *transform,
  14                                           unsigned char *src,
  15                                           unsigned char *dest,
  16                                           size_t length)
  17 {
  18     unsigned int i;
  19     float (*mat)[4] = transform->matrix;
  20     char input_back[32];
  21     /* Ensure we have a buffer that's 16 byte aligned regardless of the original
  22      * stack alignment. We can't use __attribute__((aligned(16))) or __declspec(align(32))
  23      * because they don't work on stack variables. gcc 4.4 does do the right thing
  24      * on x86 but that's too new for us right now. For more info: gcc bug #16660 */
  25     float const * input = (float*)(((uintptr_t)&input_back[16]) & ~0xf);
  26     /* share input and output locations to save having to keep the
  27      * locations in separate registers */
  28     uint32_t const * output = (uint32_t*)input;
  29
  30     /* deref *transform now to avoid it in loop */
  31     const float *igtbl_r = transform->input_gamma_table_r;
  32     const float *igtbl_g = transform->input_gamma_table_g;
  33     const float *igtbl_b = transform->input_gamma_table_b;
  34
  35     /* deref *transform now to avoid it in loop */
  36     const uint8_t *otdata_r = &transform->output_table_r->data[0];
  37     const uint8_t *otdata_g = &transform->output_table_g->data[0];
  38     const uint8_t *otdata_b = &transform->output_table_b->data[0];
  39
  40     /* input matrix values never change */
  41     const __m128 mat0  = _mm_load_ps(mat[0]);
  42     const __m128 mat1  = _mm_load_ps(mat[1]);
  43     const __m128 mat2  = _mm_load_ps(mat[2]);
  44
  45     /* these values don't change, either */
  46     const __m128 max   = _mm_load_ps(clampMaxValueX4);
  47     const __m128 min   = _mm_setzero_ps();
  48     const __m128 scale = _mm_load_ps(floatScaleX4);
  49
  50     /* working variables */
  51     __m128 vec_r, vec_g, vec_b, result;
  52
  53     /* CYA */
  54     if (!length)
  55         return;
  56
  57     /* one pixel is handled outside of the loop */
  58     length--;
  59
  60     /* setup for transforming 1st pixel */
  61     vec_r = _mm_load_ss(&igtbl_r[src[0]]);
  62     vec_g = _mm_load_ss(&igtbl_g[src[1]]);
  63     vec_b = _mm_load_ss(&igtbl_b[src[2]]);
  64     src += 3;
  65
  66     /* transform all but final pixel */
  67
  68     for (i=0; i<length; i++)
  69     {
  70         /* position values from gamma tables */
  71         vec_r = _mm_shuffle_ps(vec_r, vec_r, 0);
  72         vec_g = _mm_shuffle_ps(vec_g, vec_g, 0);
  73         vec_b = _mm_shuffle_ps(vec_b, vec_b, 0);
  74
  75         /* gamma * matrix */
  76         vec_r = _mm_mul_ps(vec_r, mat0);
  77         vec_g = _mm_mul_ps(vec_g, mat1);
  78         vec_b = _mm_mul_ps(vec_b, mat2);
  79
  80         /* crunch, crunch, crunch */
  81         vec_r  = _mm_add_ps(vec_r, _mm_add_ps(vec_g, vec_b));
  82         vec_r  = _mm_max_ps(min, vec_r);
  83         vec_r  = _mm_min_ps(max, vec_r);
  84         result = _mm_mul_ps(vec_r, scale);
  85
  86         /* store calc'd output tables indices */
  87         _mm_store_si128((__m128i*)output, _mm_cvtps_epi32(result));
  88
  89         /* load for next loop while store completes */
  90         vec_r = _mm_load_ss(&igtbl_r[src[0]]);
  91         vec_g = _mm_load_ss(&igtbl_g[src[1]]);
  92         vec_b = _mm_load_ss(&igtbl_b[src[2]]);
  93         src += 3;
  94
  95         /* use calc'd indices to output RGB values */
  96         dest[OUTPUT_R_INDEX] = otdata_r[output[0]];
  97         dest[OUTPUT_G_INDEX] = otdata_g[output[1]];
  98         dest[OUTPUT_B_INDEX] = otdata_b[output[2]];
  99         dest += RGB_OUTPUT_COMPONENTS;
 100     }
 101
 102     /* handle final (maybe only) pixel */
 103
 104     vec_r = _mm_shuffle_ps(vec_r, vec_r, 0);
 105     vec_g = _mm_shuffle_ps(vec_g, vec_g, 0);
 106     vec_b = _mm_shuffle_ps(vec_b, vec_b, 0);
 107
 108     vec_r = _mm_mul_ps(vec_r, mat0);
 109     vec_g = _mm_mul_ps(vec_g, mat1);
 110     vec_b = _mm_mul_ps(vec_b, mat2);
 111
 112     vec_r  = _mm_add_ps(vec_r, _mm_add_ps(vec_g, vec_b));
 113     vec_r  = _mm_max_ps(min, vec_r);
 114     vec_r  = _mm_min_ps(max, vec_r);
 115     result = _mm_mul_ps(vec_r, scale);
 116
 117     _mm_store_si128((__m128i*)output, _mm_cvtps_epi32(result));
 118
 119     dest[OUTPUT_R_INDEX] = otdata_r[output[0]];
 120     dest[OUTPUT_G_INDEX] = otdata_g[output[1]];
 121     dest[OUTPUT_B_INDEX] = otdata_b[output[2]];
 122 }
 123
 124 void qcms_transform_data_rgba_out_lut_sse2(qcms_transform *transform,
 125                                            unsigned char *src,
 126                                            unsigned char *dest,
 127                                            size_t length)
 128 {
 129     unsigned int i;
 130     float (*mat)[4] = transform->matrix;
 131     char input_back[32];
 132     /* Ensure we have a buffer that's 16 byte aligned regardless of the original
 133      * stack alignment. We can't use __attribute__((aligned(16))) or __declspec(align(32))
 134      * because they don't work on stack variables. gcc 4.4 does do the right thing
 135      * on x86 but that's too new for us right now. For more info: gcc bug #16660 */
 136     float const * input = (float*)(((uintptr_t)&input_back[16]) & ~0xf);
 137     /* share input and output locations to save having to keep the
 138      * locations in separate registers */
 139     uint32_t const * output = (uint32_t*)input;
 140
 141     /* deref *transform now to avoid it in loop */
 142     const float *igtbl_r = transform->input_gamma_table_r;
 143     const float *igtbl_g = transform->input_gamma_table_g;
 144     const float *igtbl_b = transform->input_gamma_table_b;
 145
 146     /* deref *transform now to avoid it in loop */
 147     const uint8_t *otdata_r = &transform->output_table_r->data[0];
 148     const uint8_t *otdata_g = &transform->output_table_g->data[0];
 149     const uint8_t *otdata_b = &transform->output_table_b->data[0];
 150
 151     /* input matrix values never change */
 152     const __m128 mat0  = _mm_load_ps(mat[0]);
 153     const __m128 mat1  = _mm_load_ps(mat[1]);
 154     const __m128 mat2  = _mm_load_ps(mat[2]);
 155
 156     /* these values don't change, either */
 157     const __m128 max   = _mm_load_ps(clampMaxValueX4);
 158     const __m128 min   = _mm_setzero_ps();
 159     const __m128 scale = _mm_load_ps(floatScaleX4);
 160
 161     /* working variables */
 162     __m128 vec_r, vec_g, vec_b, result;
 163     unsigned char alpha;
 164
 165     /* CYA */
 166     if (!length)
 167         return;
 168
 169     /* one pixel is handled outside of the loop */
 170     length--;
 171
 172     /* setup for transforming 1st pixel */
 173     vec_r = _mm_load_ss(&igtbl_r[src[0]]);
 174     vec_g = _mm_load_ss(&igtbl_g[src[1]]);
 175     vec_b = _mm_load_ss(&igtbl_b[src[2]]);
 176     alpha = src[3];
 177     src += 4;
 178
 179     /* transform all but final pixel */
 180
 181     for (i=0; i<length; i++)
 182     {
 183         /* position values from gamma tables */
 184         vec_r = _mm_shuffle_ps(vec_r, vec_r, 0);
 185         vec_g = _mm_shuffle_ps(vec_g, vec_g, 0);
 186         vec_b = _mm_shuffle_ps(vec_b, vec_b, 0);
 187
 188         /* gamma * matrix */
 189         vec_r = _mm_mul_ps(vec_r, mat0);
 190         vec_g = _mm_mul_ps(vec_g, mat1);
 191         vec_b = _mm_mul_ps(vec_b, mat2);
 192
 193         /* store alpha for this pixel; load alpha for next */
 194         dest[OUTPUT_A_INDEX] = alpha;
 195         alpha   = src[3];
 196
 197         /* crunch, crunch, crunch */
 198         vec_r  = _mm_add_ps(vec_r, _mm_add_ps(vec_g, vec_b));
 199         vec_r  = _mm_max_ps(min, vec_r);
 200         vec_r  = _mm_min_ps(max, vec_r);
 201         result = _mm_mul_ps(vec_r, scale);
 202
 203         /* store calc'd output tables indices */
 204         _mm_store_si128((__m128i*)output, _mm_cvtps_epi32(result));
 205
 206         /* load gamma values for next loop while store completes */
 207         vec_r = _mm_load_ss(&igtbl_r[src[0]]);
 208         vec_g = _mm_load_ss(&igtbl_g[src[1]]);
 209         vec_b = _mm_load_ss(&igtbl_b[src[2]]);
 210         src += 4;
 211
 212         /* use calc'd indices to output RGB values */
 213         dest[OUTPUT_R_INDEX] = otdata_r[output[0]];
 214         dest[OUTPUT_G_INDEX] = otdata_g[output[1]];
 215         dest[OUTPUT_B_INDEX] = otdata_b[output[2]];
 216         dest += RGBA_OUTPUT_COMPONENTS;
 217     }
 218
 219     /* handle final (maybe only) pixel */
 220
 221     vec_r = _mm_shuffle_ps(vec_r, vec_r, 0);
 222     vec_g = _mm_shuffle_ps(vec_g, vec_g, 0);
 223     vec_b = _mm_shuffle_ps(vec_b, vec_b, 0);
 224
 225     vec_r = _mm_mul_ps(vec_r, mat0);
 226     vec_g = _mm_mul_ps(vec_g, mat1);
 227     vec_b = _mm_mul_ps(vec_b, mat2);
 228
 229     dest[OUTPUT_A_INDEX] = alpha;
 230
 231     vec_r  = _mm_add_ps(vec_r, _mm_add_ps(vec_g, vec_b));
 232     vec_r  = _mm_max_ps(min, vec_r);
 233     vec_r  = _mm_min_ps(max, vec_r);
 234     result = _mm_mul_ps(vec_r, scale);
 235
 236     _mm_store_si128((__m128i*)output, _mm_cvtps_epi32(result));
 237
 238     dest[OUTPUT_R_INDEX] = otdata_r[output[0]];
 239     dest[OUTPUT_G_INDEX] = otdata_g[output[1]];
 240     dest[OUTPUT_B_INDEX] = otdata_b[output[2]];
 241 }
 242
 243